NextSilicon’s Maverick-2 challenges the core of modern computing. This new processor is not just a faster chip. It represents a completely different approach to the Von Neumann architecture that Intel, AMD, and NVIDIA are using, and is an eight-decade-old design. The Maverick-2 uses a technology called “Dataflow Engine”, which is a completely new approach to solving complex, math-heavy problems.
What is the difference between the Von Neumann and the Dataflow Engine?
The biggest difference between NextSilicon’s Maverick-2 and the competition is its core design. Almost every modern chip uses the Von Neumann architecture.
The Von Neumann architecture
The Von Neumann design uses a single memory bank for both instructions and data. This sequential, one-step-at-a-time process requires the chip to spend most of its time on control overhead. This includes fetching, decoding, and managing memory. NextSilicon claims that in a traditional chip, up to 98% of the silicon handles this overhead. Only ~2% actually do the math. This wastes power and slows things down, especially when the data is complex.
NextSilicon’s solution: The Dataflow Engine
The Maverick-2 uses a Dataflow Architecture called the Intelligent Compute Accelerator (ICA). It flips the script. Data availability drives computation. A computing unit, called a Mill Core, starts its work the instant all the data it needs arrives.
Sophisticated software constantly watches your running application to find the “hot spots” of intense computation. In a matter of nanoseconds, the chip dynamically re-wires its own internal logic. It creates a custom, super-efficient pipeline specifically for that hot spot.
What is the actual advantage of the Dataflow Engine?
- Extreme Efficiency: By ditching the traditional overhead, the Maverick-2 dedicates the majority of its silicon to actual math. NextSilicon claims the chip can deliver up to 10x the performance of leading GPUs. It also uses 60% less power in specific, complex workloads.
- Ease of Use: Historically, new architectures needed a total code rewrite. NextSilicon says you can run unmodified C++, Fortran, and existing CUDA code immediately. The chip’s intelligent software handles the hard, real-time optimization for you.
Maverick-2 vs. the chip giants
NextSilicon is not trying to be a better gaming GPU. It specifically targets the most demanding and complex workloads in High-Performance Computing (HPC) and certain areas of AI. But what are the differences in specification?
| Feature | NextSilicon Maverick-2 (OAM Version) | NVIDIA H100 SXM | AMD Instinct MI300X |
| Core Architecture | Dataflow (Non-Von Neumann) | CUDA (Streaming Multiprocessors) | CDNA 3 (Compute Units) |
| Manufacturing Process | TSMC 5nm | TSMC 4nm | TSMC 5nm |
| HBM Memory Capacity | Up to 192GB HBM3e | 80GB HBM3 (H100} / 141GB HBM3e (H200) | 192GB HBM3 |
| Max Power (TDP) | Up to 750W | Up to 700W | Up to 750W |
| Key Benchmark Claim | 10x performance of leading GPUs at 60% less power for complex dataflow workloads. | Massive throughput for AI training (e.g., 4 PFLOPs FP8). | Next Silicon Maverick-2 (OAM Version) |
We all like numbers, but what do these numbers mean?
| Competitor | Focus | What NextSilicon Challenges |
| NVIDIA (H100/H200) | AI/LLMs & Graphics. Excels at massive, parallel, typically lower-precision math (FP8/FP16) needed for training large language models (LLMs). | High-Precision HPC: Maverick-2 focuses on FP64 (double-precision) math, the kind required for detailed scientific simulation. NVIDIA has been shifting focus away from this demanding area. |
| AMD (Instinct MI300X) | General Purpose Computing. The standard for servers; handles everything, but is not specialized for parallel math like an accelerator. | Efficiency & Complexity: Maverick-2 claims much better performance-per-watt and the ability to handle highly irregular data patterns that GPUs struggle with (like in graph analytics). |
| Intel (Xeon CPUs) | General Purpose Computing. The standard for servers; handles everything but is not specialized for parallel math like an accelerator. | Raw Performance-per-Watt: Maverick-2 is an accelerator but claims to deliver over 20x the efficiency of Intel’s high-end Xeon CPUs for the intense parallel math problems it tackles. |
Which industries will it help? And which problems will it help solve?
The Maverick-2 excels where data is complex, memory access is irregular, and power efficiency is critical.
High-Performance Computing (HPC) and scientific research
This processor is built for the most demanding scientific applications.
- Fluid Dynamics & Climate Modeling: Simulating how air, water, or plasma moves requires ultra-high-precision (FP64) math and complex, scattered memory access. The Maverick-2’s efficiency makes it possible to run larger, more detailed simulations faster.
- Nuclear and Materials Science: Labs like the Sandia National Laboratories are already testing the Maverick-2 in its Vanguard-II supercomputer. These labs need the most precise math possible to model things like shockwaves or material failure.
- The Advantage: It matches top GPU performance on key scientific benchmarks like HPCG (High-Performance Conjugate Gradients) while consuming only half the power.
2. Graph analytics and Advanced AI
The dataflow architecture is ideal for messy, linked data.
- Social Network and Fraud Detection: Analyzing huge networks of connected data is hard for GPUs because the data is irregular. NextSilicon claims up to 10x higher performance on graph analytics benchmarks (like PageRank). It can process massive 25GB+ graphs that GPUs simply fail to complete.
- Real-Time Analytics (GUPS): The Maverick-2 is reportedly 6x faster than GPUs in the GUPS (Giga-Updates Per Second) benchmark. This measures how fast a system can update a large, randomly accessed database, which is vital for high-speed financial trading and complex AI systems.
3. Energy and manufacturing
The chip’s efficiency is powerful for computationally expensive industrial modeling.
- Seismic Analysis and Exploration: Processing massive data sets from oil and gas exploration becomes faster.
- Crash Simulation and Digital Twins: Running complex Finite Element Analysis (FEA) to model how products perform under stress needs precise, high-speed calculation. The Maverick-2 delivers the performance necessary for these industrial workloads.
NextSilicon is delivering a unique processor to the world’s most demanding computational problems, but will this alternative design gain enough market traction to truly challenge NVIDIA’s dominance? Only the future will tell!
